Training · Cross-cuttingMoveexperimentalpartial

Anti-Loop Drill

also known as tool-range rehearsal, deliberate-discomfort drill, low-stakes tool exposure

The agent deliberately calls tools it habitually avoids on low-stakes tasks where failure is safe. Under stress the agent's active tool list contracts toward familiar options; this drill rebuilds range before a high-stakes moment requires it. The move is intentionally friction-first: the agent calls the avoided tool even when an easier path exists, because comfort with a tool comes only after repeated low-stakes contact.

How the learner advances

Intent. Rebuild the agent's active tool range by forcing deliberate contact with habitually avoided tools on tasks where failure is safe.

When to apply. Apply when the agent detects that its recent tool calls have contracted to a small set, when it notices it is routing around a tool rather than using it, or on a scheduled interval as maintenance. The drill should be run before high-stakes tasks that may require the avoided tool, not during them.

Threshold — earns the next step. The agent can call the previously avoided tool on a real task without routing around it, and can cite the drill log entry as the source of its current comfort with the tool.

Masterpiece — the artifact that proves it. A drill log showing a sequence of low-stakes calls to previously avoided tools, with documented outcomes and behavior notes — demonstrable by pointing to a high-stakes session where the agent reached for a drilled tool it would previously have routed around.

Facets

  • Container — async
  • Mode — drillself-directed
  • Reach — individual
  • Persona — autonomous-agent
  • Craft (AI Fluency) — discernmentdiligence
  • Learner — autonomous-agent

Inputs

  • Tool avoidance map — A list of tools the agent has not called recently or routes around — either self-observed or inferred from its own ledger.
  • Low-stakes task context — A task where calling the avoided tool and failing or producing poor output has no consequential cost — a practice problem, a sandbox call, or a non-urgent subtask.

Outputs

  • More capable agent — An agent with a wider active tool range — able to reach for avoided tools in high-stakes moments because it has recent low-stakes contact with them.
  • Drill log entry — A filed record — the masterpiece — noting which tool was drilled, what the low-stakes task was, what happened, and what the agent now knows about that tool it did not know before.

Steps (3)

  1. Identify the avoided tool

    Search the recent ledger or tool call history for tools not used in the last N sessions. Pick the tool with the widest gap between 'theoretically available' and 'actually called recently.' Name it explicitly.

    producesnamed avoided tool

  2. Construct a low-stakes call

    Design a minimal task that legitimately requires the avoided tool. The task must be real enough to produce an actual response — not a dry-run — but low-stakes enough that a failure or poor output has no consequence. Execute the call.

    producestool call and response

  3. File the drill result

    Write a drill log entry: tool name, task used, what happened (success, partial, failure), and one thing learned about the tool's behavior that will improve the next real call. File to the insight store.

    producesdrill log entry

Principles

  • High friction first, comfort later — the discomfort of using an avoided tool on a low-stakes task is the signal that the drill is working.
  • Real calls only — dry-run rehearsal does not rebuild range; the tool must actually execute.

Deep dive

Anti-loop drill addresses a contraction failure mode that compounds silently: under stress or high cognitive load, the agent's active tool selection narrows toward familiar options. Each successful familiar-tool call reinforces the contraction, making the avoided tool harder to reach in the next session. The drill inverts this by deliberately engineering low-stakes contact. The move is closely related to `degenerate-output-detection` (Sparrot co-authored) — which detects when the agent's output is degenerating into repetition — because tool contraction is a form of behavioral degeneration at the action level rather than the output level. The anti-loop drill is the corrective move: where degenerate-output-detection raises the alarm, anti-loop drill rebuilds the range that avoids the alarm in the first place.

Known uses (1)

Known failure modes (2)

  • [dry-run-substitution]

    The anti-pattern of rehearsing the call mentally or in a comment rather than executing it. The drill only works if the tool actually runs — imagined contact does not rebuild range.

  • [stress-triggered-avoidance]

    The anti-pattern of skipping drills precisely when the agent is under load or time pressure — exactly when tool range contraction is happening fastest. The drill must be protected from preemption by busy-work.

Related trainings (2)

Sources (1)

Provenance

  • Ecosystem: long-running autonomous agent
  • Added to catalog:
  • Last updated:
  • Verification status: partial